/
rnn.py
246 lines (213 loc) · 8.48 KB
/
rnn.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
#!/usr/bin/env python3 -u
"""Time Recurrent Neural Network (RNN) for classification."""
__author__ = ["mloning"]
__all__ = ["SimpleRNNClassifier"]
from copy import deepcopy
from sklearn.utils import check_random_state
from sktime.classification.deep_learning.base import BaseDeepClassifier
from sktime.networks.rnn import RNNNetwork
from sktime.utils.validation._dependencies import _check_dl_dependencies
class SimpleRNNClassifier(BaseDeepClassifier):
"""Simple recurrent neural network.
Parameters
----------
n_epochs : int, default = 100
the number of epochs to train the model
batch_size : int, default = 1
the number of samples per gradient update.
units : int, default = 6
number of units in the network
callbacks : list of tf.keras.callbacks.Callback objects, default = None
add_default_callback : bool, default = True
whether to add default callback
random_state : int or None, default=0
Seed for random number generation.
verbose : boolean, default = False
whether to output extra information
loss : string, default="mean_squared_error"
fit parameter for the keras model
metrics : list of strings, default=["accuracy"]
metrics to use in fitting the neural network
activation : string or a tf callable, default="sigmoid"
Activation function used in the output layer.
List of available activation functions: https://keras.io/api/layers/activations/
use_bias : boolean, default = True
whether the layer uses a bias vector.
optimizer : keras.optimizers object, default = RMSprop(lr=0.001)
specify the optimizer and the learning rate to be used.
References
----------
..[1] benchmark forecaster in M4 forecasting competition:
https://github.com/Mcompetitions/M4-methods
Examples
--------
>>> from sktime.classification.deep_learning.rnn import SimpleRNNClassifier
>>> from sktime.datasets import load_unit_test
>>> X_train, y_train = load_unit_test(split="train")
>>> clf = SimpleRNNClassifier(n_epochs=20,batch_size=20) # doctest: +SKIP
>>> clf.fit(X_train, y_train) # doctest: +SKIP
ResNetClassifier(...)
"""
_tags = {
# packaging info
# --------------
"authors": ["mloning"],
# estimator type handled by parent class
}
def __init__(
self,
n_epochs=100,
batch_size=1,
units=6,
callbacks=None,
add_default_callback=True,
random_state=0,
verbose=False,
loss="mean_squared_error",
metrics=None,
activation="sigmoid",
use_bias=True,
optimizer=None,
):
_check_dl_dependencies(severity="error")
self.batch_size = batch_size
self.n_epochs = n_epochs
self.verbose = verbose
self.units = units
self.callbacks = callbacks
self.add_default_callback = add_default_callback
self.random_state = random_state
self.loss = loss
self.metrics = metrics
self.activation = activation
self.use_bias = use_bias
self.optimizer = optimizer
super().__init__()
self.history = None
self._network = RNNNetwork(random_state=random_state, units=units)
def build_model(self, input_shape, n_classes, **kwargs):
"""Construct a compiled, un-trained, keras model that is ready for training.
In sktime, time series are stored in numpy arrays of shape (d,m), where d
is the number of dimensions, m is the series length. Keras/tensorflow assume
data is in shape (m,d). This method also assumes (m,d). Transpose should
happen in fit.
Parameters
----------
input_shape : tuple
The shape of the data fed into the input layer, should be (m,d)
n_classes: int
The number of classes, which becomes the size of the output layer
Returns
-------
output : a compiled Keras Model
"""
import tensorflow as tf
from tensorflow import keras
tf.random.set_seed(self.random_state)
metrics = self.metrics if self.metrics is not None else ["accuracy"]
input_layer, output_layer = self._network.build_network(input_shape, **kwargs)
output_layer = keras.layers.Dense(
units=n_classes, activation=self.activation, use_bias=self.use_bias
)(output_layer)
self.optimizer_ = (
keras.optimizers.RMSprop(learning_rate=0.001)
if self.optimizer is None
else self.optimizer
)
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
model.compile(loss=self.loss, optimizer=self.optimizer_, metrics=metrics)
return model
def _fit(self, X, y):
"""Fit the classifier on the training set (X, y).
Parameters
----------
X : a nested pd.Dataframe, or (if input_checks=False) array-like of
shape = (n_instances, series_length, n_dimensions)
The training input samples. If a 2D array-like is passed,
n_dimensions is assumed to be 1.
y : array-like, shape = [n_instances]
The training data class labels.
Returns
-------
self : object
"""
from tensorflow import keras
y_onehot = self.convert_y_to_keras(y)
X = X.transpose(0, 2, 1)
check_random_state(self.random_state)
self.input_shape = X.shape[1:]
self.model_ = self.build_model(self.input_shape, self.n_classes_)
if self.verbose:
self.model_.summary()
# add a ReduceLROnPlateau callback is default is enabled
# if an instance of ReduceLROnPlateau is already present
# then don't add it again.
if self.add_default_callback:
reduce_lr = keras.callbacks.ReduceLROnPlateau(
monitor="loss",
factor=0.5,
patience=50,
min_lr=0.0001,
)
if self.callbacks is None:
self.callbacks_ = [
reduce_lr,
]
elif isinstance(self.callbacks, keras.callbacks.Callback):
self.callbacks_ = [
self.callbacks,
reduce_lr,
]
elif isinstance(self.callbacks, tuple):
self.callbacks_ = deepcopy([i for i in self.callbacks])
if not any(
isinstance(callback, keras.callbacks.ReduceLROnPlateau)
for callback in self.callbacks
):
self.callbacks_.append(reduce_lr)
else:
raise TypeError(
"`callback` can either be None, an instance "
"of keras.callbacks.Callback or a tuple containing "
"keras.callbacks.Callback objects. "
f"But found {type(self.callbacks)} instead."
)
else:
self.callbacks_ = deepcopy(self.callbacks)
self.history = self.model_.fit(
X,
y_onehot,
batch_size=self.batch_size,
epochs=self.n_epochs,
verbose=self.verbose,
callbacks=self.callbacks_,
)
return self
@classmethod
def get_test_params(cls, parameter_set="default"):
"""Return testing parameter settings for the estimator.
Parameters
----------
parameter_set : str, default="default"
Name of the set of test parameters to return, for use in tests. If no
special parameters are defined for a value, will return ``"default"`` set.
Reserved values for classifiers:
"results_comparison" - used for identity testing in some classifiers
should contain parameter settings comparable to "TSC bakeoff"
Returns
-------
params : dict or list of dict, default = {}
Parameters to create testing instances of the class
Each dict are parameters to construct an "interesting" test instance, i.e.,
``MyClass(**params)`` or ``MyClass(**params[i])`` creates a valid test
instance.
``create_test_instance`` uses the first (or only) dictionary in ``params``
"""
params1 = {}
params2 = {
"n_epochs": 50,
"batch_size": 2,
"units": 5,
"use_bias": False,
}
return [params1, params2]